Interestingly, this recent study using ChatGPT Health gave quite a different outcome (https://www.nature.com/articles/s41591-026-04297-7). Here it was wrong about emergency triage 50% of the time.